Pushing memory bound kernels beyond the speed of light with lossless decompression

The fastest a memory bound kernel can go is set by the time required to transfer the data to the SMs. How can we do better?